Unverified50% confidenceFactExact time
FP8是NVIDIA Blackwell架构原生支持的数据类型，相比INT量化保留了浮点数的动态范围表达能力
1
Sources
50%
Confidence
Long-term
Relevance
8/7/2026
First Seen
Sources
Related Claims
UnverifiedFP8是NVIDIA Blackwell架构原生支持的数据类型，分为E4M3和E5M2两种格式80% similarVerifiedFP8是NVIDIA Hopper架构（H100）原生支持的8-bit浮点格式，相比INT8量化保留了更好的动态范围77% similarUnverifiedFP8是Blackwell架构原生支持的低精度计算格式，包含E4M3和E5M2两种变体，Tensor Core的FP8计算吞吐是FP16的两倍73% similarUnverifiedBlackwell架构引入第五代张量核心，原生支持FP4×FP4矩阵乘加运算，累加器使用FP32或FP16精度69% similarUnverifiedBlackwell是NVIDIA继Hopper和Ada Lovelace之后的新一代微架构，引入第五代Tensor Core以提升FP4/FP6推理效率67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/700139API
curl https://kongchang.com/api/v1/knowledge/claims/700139MCP
get_claim(id=700139)